Musculoskeletal disease risk assessment and early warning method and system based on multi-source data

Abstract

The invention provides a musculoskeletal disease risk assessment and early warning method and system based on multi-source data, and relates to the technical field of assessment and early warning of muscles and skeletons. Joint angles, myoelectricity and plantar load data are collected, a joint feature vector sequence consistent in time sequence is constructed, the limitation of a traditional single monitoring means is effectively made up, and the risk assessment and early warning of musculoskeletal diseases is achieved. K-means clustering is carried out on a joint angle sequence to form a standard attitude set with physiological significance, and a transition edge weight based on angle change, muscle activation and load fluctuation weighted calculation is constructed, so that quantitative expression of physiological costs of different attitude transfer paths is realized, a risk transition region is conveniently identified, and the risk transition rate is improved. According to the method, the sudden transition behavior in the movement process can be captured by calculating the deviation between the predicted value and the true value of the latent variable, and effective early warning of the musculoskeletal disease risk is realized by combining the edge weight strength of the sudden transition behavior in the directed graph and considering the transition strength and the abnormal behavior degree at the same time.

Description

Technical Field

[0001] The present invention relates to the technical field of muscle and bone assessment and early warning, and specifically provides a method and system for risk assessment and early warning of musculoskeletal disorders based on multi-source data. Background Art

[0002] In recent years, with the development of intelligent rehabilitation, sports health management, and industrial ergonomics fields, higher requirements have been put forward for the dynamic monitoring and risk early warning of key joints in the lower limbs of the human body. These joints bear high-frequency physiological load changes during human movement and are the high-incidence sites of many musculoskeletal disorders. Especially in high-intensity labor, rehabilitation training, or sports competition scenarios, individuals are extremely prone to irreversible injuries due to sudden movements, poor postures, or fatigue accumulation. Therefore, early detection and early warning of the complex physiological loads borne by the lower limb joints of the human body during movement, labor, and rehabilitation, and the real-time assessment of their movement states are of great significance for preventing musculoskeletal injuries.

[0003] Common methods in the prior art include using inertial measurement units for joint angle monitoring, using surface electromyography to reflect muscle activation states, or estimating gait load through plantar pressure sensors. These methods can provide certain physiological parameter information in their respective dimensions and are widely used in clinical rehabilitation and sports science research. However, they are often used in isolation, lacking fusion and correlation modeling between data, and it is difficult to form an overall portrait of movement behavior.

[0004] The prior art still has obvious deficiencies in many aspects: First, single-dimensional signals cannot comprehensively depict the complex evolution process of joint states, especially in complex dynamic movements, and cannot effectively reflect the coupling relationship between different physiological variables; Second, existing methods generally lack the ability to model the "transition" process of movement states (such as sudden postures, abnormal movements), resulting in a lag or even omission in the identification of potential abnormalities.

[0005] The above information disclosed in the background art section is only used to enhance the understanding of the background of the present disclosure, and thus it may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0006] The purpose of the present invention is to provide a method and system for risk assessment and early warning of musculoskeletal disorders based on multi-source data to solve the problems raised in the above background art.

[0007] To achieve the above purpose, the present invention provides the following technical solutions: A method for risk assessment and early warning of musculoskeletal disorders based on multi-source data, the specific steps including: Continuously collect multi-channel data of the target joint to form a joint feature vector sequence. The multi-channel data includes joint angle data corresponding to time stamps, electromyogram signal data of the muscle surface associated with the target joint, and corresponding joint load data. The joint feature vector sequence includes a joint angle sequence, an electromyogram signal sequence, and a joint load sequence; Perform K-means clustering on the joint angle sequence, divide all joint angle data into multiple standard postures, construct the edge weights between any two standard postures according to the joint angle data, electromyogram signal data, and corresponding joint load data within each standard posture, and construct a directed graph network representing the behavior evolution path according to the acquisition order; Input the joint feature vector sequence into a variational autoencoder model for feature compression to obtain a low-dimensional latent variable representation sequence. Input the latent variable representation sequence into a time series prediction model, use a gated recurrent unit for time series modeling, and predict the latent variable representation of the next sampling point based on the latent variable representation of the previous sampling point in the latent variable representation sequence; Calculate the deviation between each actual latent variable and the predicted latent variable, determine whether there is an abnormal transition behavior between the multi-channel data corresponding to adjacent sampling points according to the deviation, map the abnormal transition behavior into the directed graph network, count the cumulative frequency and edge weight distribution of the abnormal transition behavior, and divide the risk level, and output the corresponding musculoskeletal disorder warning signal.

[0008] Further, the target joint is one of the hip joint, knee joint, or ankle joint in the lower limb joints, and the joint load data is the average pressure on the soles of both feet; When collecting multi-channel data of the target joint, set multiple sampling points, synchronously collect joint angle data, electromyogram signal data, and joint load data at each sampling point, and the time interval between adjacent sampling points is the same.

[0009] Further, the logic of performing K-means clustering on the joint angle sequence is as follows: Determine the maximum movement angle of the target joint, divide the maximum movement angle into angle ranges corresponding to K standard postures, and use the median of the angles within each angle range as the initialized clustering center; Calculate the Euclidean distance between each joint angle data and each initialized clustering center, assign the corresponding joint angle data to the nearest clustering center, and update each clustering center to the mean of all angle values in its clustering cluster; Repeat calculating the Euclidean distance between each joint angle data and each updated clustering center, reassign the corresponding joint angle data to the nearest clustering center until the clustering center converges, that is, the clustering center no longer changes; Map the standard posture corresponding to the cluster where each joint angle data is located to the standard posture corresponding to each joint angle data.

[0010] Furthermore, the method for constructing the edge weights between any two standard postures is as follows: Calculate the average of the joint angles, the root mean square value of all EMG signals, and the average of the joint loads within the clustering clusters corresponding to the two standard postures respectively, and calculate the difference values of the joint angles, EMG signals, and joint loads between the two clustering clusters. The formulas are as follows: Where, , and respectively represent the difference values of the joint angles, EMG signals, and joint loads between the clustering clusters corresponding to the standard postures and ; and respectively represent the average of the joint angles of the corresponding clustering clusters when the standard postures are and ; and respectively represent the average of the joint loads of the corresponding clustering clusters when the standard postures are and ; and respectively represent the root mean square value of the EMG signals on the th associated muscle surface in the corresponding clustering clusters when the standard postures are ; and both represent the numbers of the standard postures; and respectively represent the number of the muscle associated with the target joint and the total number of associated muscles; For the difference values, perform weighting to construct the edge weights between the two standard postures. The formula is as follows: Where, represents the edge weight when the standard posture transitions to the standard posture ; respectively represent the angle weight, signal weight, and load weight, and , and .

[0011] Furthermore, the method for constructing a directed graph network representing the behavior evolution path is as follows: Taking the joint angle data, EMG signal data, and corresponding joint load data at each sampling point as a time node, sorting all the time nodes according to the acquisition timestamps corresponding to the acquisition points, the direction of the directed graph network is from the previous time node to the next time node, and the edge weight between two adjacent time nodes is the edge weight when transitioning from the standard pose corresponding to the previous time node to the standard pose corresponding to the next time node.

[0012] Further, the joint feature vector sequence is: Where, is the joint feature vector of the th sampling point, and respectively represent the joint angle data and joint load data of the th sampling point, represents the EMG signal data of the th associated muscle surface among the th sampling point, represents the total number of muscles associated with the target joint, represents the total number of sampling points; The method for obtaining the low-dimensional latent variable representation sequence is: Inputting the joint feature vector of each sampling point into the trained variational autoencoder, outputting the mean vector and logarithmic variance vector of the latent variable, and sampling the latent variable representation corresponding to each sampling point from the latent variable distribution based on the reparameterization technique. The formula is: Where, and respectively represent the latent variable representation and mean vector of the th sampling point, , represents the logarithmic variance vector, is the standard normal noise, that is, obeys the multi-dimensional Gaussian distribution with a mean of 0 and a covariance matrix of the identity matrix; Forming a low-dimensional latent variable representation sequence according to the timestamp order corresponding to the sampling points. The latent variable representation sequence is , and the method for predicting the latent variable representation of the next sampling point based on the latent variable representation of the previous sampling point in the latent variable representation sequence is: Using the trained gated recurrent unit to perform temporal modeling on the latent variable sequence . Taking the latent variable representation of each sampling point as the input, the initial hidden state of the gated recurrent unit is 0, receiving the latent variable representation of the current sampling point and the hidden state of the previous sampling point, and calculating the hidden state of the current sampling point through the gating mechanism; Based on the hidden state of the current sampling point, use a linear layer to predict the latent variable representation of the next sampling point. The logic is as follows: Among them, represents the predicted latent variable representation of the th sampling point, represents the hidden state of the th sampling point, and respectively represent the learned weights and biases of the gated recurrent unit.

[0013] Furthermore, calculate the deviation between the predicted latent variable representation and the true latent variable representation of each sampling point. The formula is as follows: Among them, represents the deviation between the predicted latent variable representation and the true latent variable representation of the th sampling point, and respectively represent the predicted latent variable representation and the true latent variable of the th sampling point. When , it is determined that there is an abnormal transition behavior between the multi-channel data of the th sampling point and the th sampling point, represents the deviation threshold.

[0014] Furthermore, the method for outputting the corresponding musculoskeletal disorder warning signal is as follows: Map the abnormal transition behavior to a directed graph network, obtain the edge weights corresponding to each abnormal transition behavior, count the frequency of the abnormal transition behavior, and combine the deviation corresponding to each abnormal transition behavior to calculate the transition risk state of the target joint during the observation process. The formula is as follows: Among them, represents the transition risk state, represents the total number of all sampling points, that is, it represents the total number of all edge weights, and respectively represent the edge weight and deviation corresponding to the th abnormal transition behavior, represents the total number of all abnormal transition behaviors; If , issue a low-risk warning indicating occasional abnormal transitions with low edge weights. If , issue a medium-risk warning indicating frequent abnormal transitions or significant edge weights. If , issue a high-risk warning indicating multiple abnormal transitions occurring on high-edge weight paths. If , issue an ultra-high risk warning requiring immediate intervention.

[0015] The present invention further provides a musculoskeletal disorder risk assessment and early warning system based on multi-source data. The system is used to execute the above-mentioned method for risk assessment and early warning of musculoskeletal disorders based on multi-source data, and includes: A feature acquisition module, which is used to continuously acquire multi-channel data of a target joint to form a joint feature vector sequence. The multi-channel data includes joint angle data corresponding to time stamps, electromyogram signal data on the muscle surface associated with the target joint, and corresponding joint load data. The joint feature vector sequence includes a joint angle sequence, an electromyogram signal sequence, and a joint load sequence; A network construction module, which is used to perform K-means clustering on the joint angle sequence, divide all joint angle data into multiple standard postures, construct the edge weights between any two standard postures according to the joint angle data, electromyogram signal data, and corresponding joint load data within each standard posture, and construct a directed graph network representing the behavior evolution path according to the acquisition order; A variable prediction module, which is used to input the joint feature vector sequence into a variational autoencoder model for feature compression to obtain a low-dimensional latent variable representation sequence, input the latent variable representation sequence into a time series prediction model, perform time series modeling using a gated recurrent unit, and predict the latent variable representation of the next sampling point based on the latent variable representation of the previous sampling point in the latent variable representation sequence; A risk assessment module, which is used to calculate the deviation between each actual latent variable and the predicted latent variable, determine whether there is an abnormal transition behavior between the multi-channel data corresponding to adjacent sampling points according to the deviation, map the abnormal transition behavior into the directed graph network, count the cumulative frequency and edge weight distribution of the abnormal transition behavior, divide the risk level, and output the corresponding musculoskeletal disorder early warning signal.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By collecting joint angle, electromyogram, and plantar load data, the present invention constructs a joint feature vector sequence with consistent time series, effectively making up for the limitations of traditional single monitoring means and providing a data basis for subsequent multi-modal joint modeling. Secondly, by performing K-means clustering on the joint angle sequence, a set of standard postures with physiological significance is formed, and the transition edge weights are constructed based on weighted calculation of angle change, muscle activation, and load fluctuation, realizing the quantitative expression of the physiological cost of different posture transition paths and facilitating the identification of risk transition regions.

[0017] The variational autoencoder is used to compress the high-dimensional joint features to obtain the low-dimensional expression of the behavior state in the latent variable space, and the gated recurrent unit is used to predict and model the latent variable sequence, thereby establishing a benchmark trajectory for the temporal evolution of individual behaviors. On this basis, the deviation between the predicted value and the true value of the latent variable is calculated, which can capture the sudden transition behaviors during the movement process. By combining the edge weight strength in the directed graph and considering both the transition strength (edge weight) and the behavior anomaly degree (latent variable deviation), an effective early warning of the risk of musculoskeletal disorders can be achieved. Description of the Drawings

[0018] Figure 1 Schematic diagram of the overall method flow of the present invention; Figure 2 Schematic diagram of the joint feature vector sequence in the present invention; Figure 3 Schematic diagram of the directed graph network in the present invention; Figure 4 Schematic diagram of the overall method flow of the present invention; Figure 5 Schematic diagram of the generation of the transition risk state in the present invention. Detailed Embodiments

[0019] In order to make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the following further elaborates on the present invention in conjunction with specific embodiments.

[0020] It should be noted that unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meanings understood by those of ordinary skill in the field to which the present invention belongs. The "first", "second", and similar terms used in the present invention do not indicate any order, quantity, or importance, but are only used to distinguish different components. The terms such as "including" or "comprising" mean that the elements or objects appearing before this word cover the elements or objects listed after this word and their equivalents, without excluding other elements or objects. The terms such as "connected" or "linked" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left", and "right" are only used to indicate relative position relationships, and when the absolute position of the object being described changes, the relative position relationship may also change accordingly.

[0021] Embodiment

[0022] Please refer to Figures 1 - 4 , the present invention provides a technical solution: A method for risk assessment and early warning of musculoskeletal disorders based on multi-source data, and the specific steps include: Step 1: Continuously collect multi-channel data of the target joint to form a joint feature vector sequence. The multi-channel data includes joint angle data corresponding to time stamps, electromyography signal data on the surface of the muscles associated with the target joint, and corresponding joint load data. The joint feature vector sequence includes a joint angle sequence, an electromyography signal sequence, and a joint load sequence.

[0023] In this embodiment, the target joint is one of the hip joint, knee joint, or ankle joint in the lower limb joints. The joint load data is the average pressure on the soles of both feet. The hip joint, knee joint, or ankle joint is selected as the monitoring object because, as the core weight-bearing and movement hubs of the human lower limb, they undertake key mechanical and control functions in various high-frequency activities. These three types of joints are the high-incidence sites of clinically musculoskeletal disorders (such as arthritis, meniscus injury, patellar chondromalacia, etc.), and are often strained and diseased due to repeated high loads, poor movement patterns, or muscle control disorders. Therefore, continuously monitoring the angle changes, muscle activation states, and mechanical load conditions of the key joints in the lower limb during movement is of great significance for early identification of disease risks and providing early warning prompts.

[0024] Furthermore, the joint angle data refers to the relative angle between bone segments formed by the target joint during movement. The joint angle in the knee joint is the flexion and extension angle between the thigh (femur) and the calf (tibia). The joint angle in the hip joint is the angle between the thigh and the torso, covering flexion and extension, adduction and abduction, and rotation. The joint angle in the ankle joint is the angle between the calf and the foot, covering plantar flexion (tiptoe pressing down) and dorsiflexion (tiptoe lifting) and other movements. The joint angle data is obtained by wearing an inertial measurement unit on parts such as the thigh and calf, collecting triaxial acceleration and angular velocity data, and combining existing attitude calculation algorithms (such as Madgwick, Kalman filtering) to calculate the bone segment attitude, thereby obtaining the joint angle.

[0025] The movement of the target joint is actively driven by the surrounding muscle groups. Therefore, monitoring the surface electromyography signals of the muscles associated with the target joint can indirectly reflect the neural control load, muscle coordination, and fatigue degree of the joint. The muscles associated with the knee joint include the quadriceps femoris, biceps femoris, and gastrocnemius. The muscles associated with the hip joint include the gluteus maximus, iliopsoas, and middle gluteal muscles. The muscles associated with the ankle joint include the tibialis anterior, soleus, and gastrocnemius. The surface electromyography signals are collected through surface electrode patches and obtained by connecting to an electromyography acquisition instrument.

[0026] In this embodiment, the joint load data indirectly estimates the vertical biomechanical load borne by the target joints through the average pressure on the soles of both feet. The joint loads of the lower limbs of the human body mainly come from the downward transmission of body gravity, the transmission of ground reaction force from bottom to top, and the additional impact force caused by movement acceleration and inertia. These forces are transmitted upward to the ankle, knee, hip and other joints through the soles of the feet, forming stress on the joint cartilage, ligaments and tendons. The average pressure on the soles of both feet is obtained by collecting and obtaining it through an insole pressure sensor. The insole pressure sensor has multiple built-in pressure points and can calculate the average pressure.

[0027] When collecting multi-channel data of the target joint, multiple sampling points are set, and each sampling point synchronously collects joint angle data, electromyographic signal data and joint load data, and the time intervals of adjacent sampling points are the same. In order to achieve high timeliness and synchronous monitoring of the physiological load state of the target joint, multi-channel synchronous sampling is performed at a fixed time interval. Specifically, multiple sampling points are set during the collection process. The sampling points are discrete time nodes divided by the system on the time axis according to the set sampling frequency. Each sampling point corresponds to a specific collection moment, and multiple sampling points form a complete time series in timestamp order.

[0028] At each sampling point, multi-channel data of the target joint is collected simultaneously, including: joint angle data at that moment, electromyographic signal data of multiple muscle surfaces associated with the target joint, and the average pressure value of the soles of both feet as joint load data. The sampling process has a strict time alignment characteristic, that is, the above three types of data are synchronously acquired at the same sampling point to ensure that they have consistent time labels and can accurately reflect the movement posture, muscle activation status and load of the target joint at the same moment. The time intervals between sampling points are consistent to form an equally spaced time series, so that the system can perform dynamic modeling and behavior evolution analysis based on a unified time scale. The multi-channel joint feature vector sequence constructed in this way has a high degree of temporal consistency and physiological parameter coordination, which provides a data foundation and modeling guarantee for subsequent latent variable compression, sequence prediction and anomaly detection.

[0029] Step 2: Perform K-means clustering on the joint angle sequence, divide all joint angle data into multiple categories of standard postures, construct the edge weights between any two standard postures based on the corresponding joint angle data, electromyographic signal data and corresponding joint load data within each category of standard posture, and construct a directed graph network representing the behavior evolution path according to the acquisition order.

[0030] In this embodiment, the logic of K-means clustering for joint angle sequences is: Determine the maximum range of motion of the target joint, and divide the maximum range of motion into angular ranges corresponding to K standard postures. The maximum range of motion is the difference interval between the minimum angle value and the maximum angle value that the joint may reach during normal movement. Divide this angular range into equal intervals according to the set number of clusters K to form initial angular sub-intervals corresponding to K standard postures. The median of each angular sub-interval is selected as the initial clustering center of the standard posture, constituting an initial set of K clustering centers; For the joint angle data of each sampling point, calculate the Euclidean distance between it and all K initialized clustering centers respectively, and assign the angle data to the cluster corresponding to the nearest clustering center. After one round of assignment, calculate the average value of the angle values within each cluster to obtain a new clustering center. The new clustering center is the arithmetic mean of all angle values in the current cluster, replacing the original center position; Recalculate the Euclidean distance between all angle data and the new clustering centers, and reassign the data; then update the mean value of the angle values within the new clusters. The above data assignment and clustering center update processes are executed iteratively until the change value between all clustering centers is lower than the preset convergence threshold or the positions of the clustering centers no longer change, that is, it is judged that the clustering process is completed, and finally K stable clustering centers are obtained.

[0031] After clustering is completed, the system maps each joint angle data to the corresponding standard posture number according to the category label of the cluster it belongs to. Thus, the original continuous joint angle sequence is converted into a discretized standard posture label sequence, providing a structural basis for subsequent directed graph modeling and behavior evolution analysis based on posture transitions; Specifically, to achieve the standard posture division of the target joint motion state, it is necessary to first clarify the maximum range of motion of the target joint, and perform K-means clustering on the collected joint angle data based on this angular range, so as to obtain multiple standard posture categories. The division method aims to ensure the physiological rationality and numerical balance of the posture interval division on the basis of ensuring the stability of clustering initialization, so as to improve the clustering convergence efficiency and the representativeness of the posture discretization results. Different lower limb joints of the human body have clear amplitude limits in physiological structure, and this range usually reflects the maximum flexion and extension movement trajectories that the joint can complete under active control. The system uses this maximum range of motion as the division basis, which helps to form a reasonable coverage of the complete movement amplitude in the initial stage of clustering, and avoids the initialization clustering centers gathering in a narrow angular interval, affecting the final clustering effect.

[0032] Taking the knee joint as an example, the common range of motion is from 0° (fully extended) to 135° (fully flexed). The 135° range is divided into K equal-width angular intervals. Given that the number of clusters K = 6, the width of each interval is approximately 22.5°, forming six initial posture intervals: [0°, 22.5°], [22.5°, 45°], [45°, 67.5°], [67.5°, 90°], [90°, 112.5°], and [112.5°, 135°]. The corresponding initial cluster centers are set as the mid-values of each interval, namely 11.25°, 33.75°, 56.25°, 78.75°, 101.25°, and 123.75°. The interval [0°, 22.5°] corresponds to full knee extension or a near-erect state, commonly seen in static standing or the initial preparatory posture. The interval [22.5°, 45°] corresponds to a mildly flexed knee posture, such as the stance phase during starting or slow walking. The interval [45°, 67.5°] corresponds to a moderately flexed knee posture, commonly seen during stair climbing or the transitional phase of movement. The interval [67.5°, 90°] corresponds to a standard squat posture or a working squat state. The interval [90°, 112.5°] corresponds to the deep knee flexion phase, such as static load-bearing actions like sitting on the ground or half-kneeling. The interval [112.5°, 135°] corresponds to the extreme knee flexion state, often occurring during ground-assisted actions or the instant of returning to a standing position from a squat.

[0033] If the target joint is the hip joint, its common maximum range of motion is from -30° to 120°, corresponding to a total flexion-extension range of 150°. Under the clustering condition of K = 5, the width of each angular interval is 30°, divided into [-30°, 0°], [0°, 30°], [30°, 60°], [60°, 90°], and [90°, 120°]. The cluster centers are -15°, 15°, 45°, 75°, and 105° respectively, covering the entire process from maximum extension to maximum flexion. The interval [-30°, 0°] indicates that the hip joint is in a hyperextended state, commonly seen in fast walking or leg backward swing. The interval [0°, 30°] indicates that the hip joint is in an erect or slightly flexed state, typical of standing and gait support. The interval [30°, 60°] indicates moderate hip flexion, such as in the middle of walking or taking a large step. The interval [60°, 90°] indicates deep hip flexion, such as sitting down or lifting the leg when going upstairs. The interval [90°, 120°] indicates the extreme hip flexion state, such as actions like hugging the knees or being in a deep sitting position.

[0034] For the ankle joint, the range of dorsiflexion and plantarflexion is generally from -20° to 45°, with a total of approximately 65°. It can be divided into equal-width intervals under the condition of K = 4, and each posture covers approximately 16.25°. The corresponding clustering centers are the mid-values of each interval, such as -10°, 8.125°, 24.375°, and 38.75°. The [-20°, 0°] interval indicates that the ankle joint is in the dorsiflexion state, which is common when the foot swings off the ground or during landing cushioning. The [0°, 16.25°] interval represents the neutral or mild plantarflexion state, seen in static standing or slow walking. The [16.25°, 32.5°] interval indicates moderate plantarflexion, such as when stepping on the ground before takeoff or during the extension phase. The [32.5°, 45°] interval represents strong plantarflexion, which often appears in strenuous movements such as jumping off the ground and toe extension.

[0035] Furthermore, the method for constructing the edge weights between any two standard postures is as follows: Calculate the average of the joint angles, the root mean square value of all electromyography signals, and the average of the joint loads within the clustering clusters corresponding to the two standard postures respectively, and calculate the difference values of the joint angles, electromyography signals, and joint loads between the two clustering clusters. The formulas are as follows: Among them, 、 and represent the difference values of the joint angles, electromyography signals, and joint loads between the clustering clusters corresponding to the standard postures and respectively. and represent the average joint angles of the clustering clusters corresponding to the standard postures and respectively. and represent the average joint loads of the clustering clusters corresponding to the standard postures and respectively. and represent the root mean square values of the electromyography signals on the -th and -th associated muscle surfaces of the clustering clusters corresponding to the standard postures and and both represent the numbers of the standard postures. and represent the numbers of the muscles associated with the target joint and the total number of associated muscles respectively.

[0036] Furthermore, in order to quantitatively describe the transition behavior between standard postures, it is necessary to construct the edge weights between any two standard postures. The edge weights are used to measure the state transition intensity between two standard postures, comprehensively reflecting the degree of angular change of the target joint, the activation difference of the involved muscles, and the fluctuation of the joint load during the transition process. As an attribute of the connection path between standard postures, the edge weights are an important basis for subsequent construction of a directed graph network structure, identification of abnormal transition behaviors, and individual risk scoring.

[0037] The process of constructing edge weights depends on the statistical analysis of the data in each standard posture clustering cluster. Specifically, first, for two standard postures to be compared and , the data samples of all sampling points within their respective clustering clusters are extracted, and based on these samples, the average values of three key physiological indicators are calculated: namely, the average value of joint angles, the root mean square value of electromyography of the muscle groups associated with the target joint, and the average sole pressure value as the average value of joint load. Among them, the average joint angle represents the amplitude of the movement corresponding to this posture, the average joint load reflects the mechanical load at the steady state of this posture, and the root mean square value of electromyography is used to represent the average activation degree of the muscles mainly involved in control and stability in this posture. Assume that there are G sampling points in the clustering cluster of the standard posture , and the signal of the kth associated muscle at the th sampling point is , then the root mean square value of this muscle signal under the standard posture is , and the calculation formula is: .

[0038] In this embodiment, by calculating the difference values of joint angles, electromyography signals, and joint loads, the physiological state changes between standard postures can be comprehensively characterized, and further, the movement amplitude changes, physiological energy consumption changes, and mechanical load changes of an individual during action transition can be reflected. The magnitude of the above difference values directly reflects the complexity of the action transition and the potential risk impact on the musculoskeletal system.

[0039] The joint angle difference value reflects the degree of change in the range of joint motion during the transition from one standard posture to another. The larger the difference value, the more drastic the change in the range of joint motion during the transition. The joint needs to quickly complete complex movements such as large-scale flexion and extension or rotation, which may lead to increased soft tissue traction and joint stress concentration, thereby increasing the fatigue risk and injury probability of the musculoskeletal system. The electromyographic signal difference value is used to reflect the activation changes of the target joint-related muscles during the action transition. The root mean square value of the electromyographic signal is an effective indicator for measuring the intensity of muscle activity. The larger the difference value, the more the muscle needs to significantly increase or decrease the contraction intensity to cooperate with the joint to complete the posture change during the transition. This large-scale muscle activation change is often accompanied by higher energy consumption and muscle fatigue accumulation. It may also reflect that the individual has muscle control imbalance or compensation during the transition, which is an important basis for evaluating muscle load changes and fatigue risks.

[0040] The difference value of joint load directly reflects the change of external mechanical load borne by the joints of the individual during the jump process. The joint load is usually reflected by the average pressure of the soles of both feet. The larger the difference value, the more dramatic the change of ground reaction force or body weight load during the action, which may cause the instantaneous force on the joint to increase sharply. In this case, the mechanical impact on the joint cartilage, ligaments and tendons increases, further exacerbating the wear and damage risk of joint tissues. The larger the difference value of joint angle, electromyographic signal and joint load, the more dramatic the change of exercise intensity, muscle load and mechanical stress experienced by the individual when jumping from the current standard posture to the target standard posture. The jump process corresponding to the large difference value often has a higher physiological burden and potential injury risk, which is a high-risk moment for the occurrence of musculoskeletal diseases. Therefore, the edge weight calculated based on these difference values ​​can effectively characterize the comprehensive risk level of the jump process.

[0041] The difference values ​​are weighted to construct the edge weight between the two standard postures. The formula is: in, Indicates standard posture Towards standard posture The edge weight when making a transition, denote angle weight, signal weight and load weight respectively, ,and .

[0042] Furthermore, the edge weights between standard postures are designed to quantify the transition intensity from one standard posture to another, as well as the impacts of joint movements, muscle activations, and load changes involved in this transition process on an individual's physiological system. By taking the weighted average of joint angle differences, electromyography signal differences, and joint load difference values, the transition intensity and physiological load level between standard postures can be scientifically measured, thereby assigning a numerical "edge weight" to each edge in the constructed directed graph network for subsequent abnormal transition analysis, risk assessment, and warning mechanisms. The weights respectively represent the importance weights of joint angle differences, electromyography signal differences, and joint load differences in edge weight calculation. The change in joint angle directly reflects the change in movement amplitude. A larger angle change usually means that the joint needs to bear a higher physiological burden. Therefore, the angle difference accounts for a relatively high weight in the edge weight. It is set that indicates that the angle change has a higher priority when evaluating the transition intensity. The change in electromyography signal reflects the change in muscle activation degree and muscle control ability. A larger electromyography signal difference usually means a heavier muscle burden and can also reflect possible muscle fatigue or imbalance during the movement process. Therefore, it has a certain weight in the edge weight. The difference in joint load reflects the change in mechanical load borne by the joint. Although the load difference is very important in some actions, compared with the changes in angle and electromyography, the change in joint load is usually more affected by the movement mode and external environment. Therefore, the weight of the load difference is relatively small.

[0043] According to actual applications and data analysis, the weights can be adjusted according to the characteristics of the target joint and the needs of the research object. In this embodiment, it can be set to . This setting can ensure that the change in joint angle dominates, and the changes in electromyography signal and load moderately affect the final edge weight calculation, which is applicable to most physiological activity monitoring scenarios.

[0044] The magnitude of the edge weight value directly reflects the physiological load degree of the transition between standard postures. The larger the edge weight, the greater the change in joint angle during the transition process, the larger the movement amplitude, the larger the change in electromyography signal, the higher the muscle activation degree, and there may be a greater risk of fatigue accumulation or imbalance. The change in joint load is larger, indicating that the mechanical load borne by the joint changes sharply during the transition process, which may increase the risk of joint injury.

[0045] Furthermore, the method for constructing a directed graph network representing the behavior evolution path is as follows: Taking the joint angle data, EMG signal data, and corresponding joint load data at each sampling point as a time node, all time nodes are sorted according to the acquisition timestamp corresponding to the acquisition point. The direction of the directed graph network is from the previous time node to the next time node, and the edge weight between two adjacent time nodes is the edge weight when transitioning from the standard pose corresponding to the previous time node to the standard pose corresponding to the next time node.

[0046] To accurately represent the behavior evolution path of the target joint during movement and construct a directed graph network based on the transition of standard poses, the system first combines and processes the joint angle data, EMG signal data, and joint load data at each sampling point. The data at each sampling point is a time node, representing the state of the target joint at a specific moment. These time nodes are sorted according to the timestamp order of the acquisition points, forming a strictly chronological time series. In this time series, the data at each sampling point constitutes a node in the graph, and this node contains the joint angle, EMG signal intensity, and joint load data of the target joint at that moment. To form a directed graph network, directed edges in the graph are established in the order from the previous time node to the next time node. Each edge represents the transition process from the standard pose to the standard pose. The edge weight in the graph reflects the "physiological transition intensity" between two adjacent time nodes, that is, the edge weight value from the standard pose corresponding to the previous time node to the standard pose corresponding to the next time node.

[0047] In this directed graph network, the time nodes represent the pose states of the individual at each time node, and the edges represent the transition process from one pose to another. The edge weight value between two adjacent time nodes is calculated by weighting the difference values of angles, EMGs, and loads from the previous standard pose to the next standard pose. This edge weight value reflects the amplitude of the target joint pose change, the change in muscle activation, and the degree of joint load change during a specific time period. A larger edge weight value indicates that during this transition process, the target joint has experienced a larger range of motion or load change, with a heavier physiological burden and potentially higher risks. In this way, the movement process of the target joint is transformed into a directed graph network, and the edge weights between nodes quantify the individual's behavior evolution trajectory through the transition of poses during the movement process. This directed graph network provides a basis for subsequent abnormal transition detection, risk assessment, and output of warning signals, and can timely identify and accurately feedback potential risks during the movement process.

[0048] Step 3: Input the joint feature vector sequence into the variational autoencoder model for feature compression to obtain a low-dimensional latent variable representation sequence. Input the latent variable representation sequence into the time series prediction model, use the gated recurrent unit for time series modeling, and predict the latent variable representation of the next sampling point based on the latent variable representation of the previous sampling point in the latent variable representation sequence.

[0049] In this embodiment, the joint feature vector sequence is: where is the joint feature vector of the th sampling point, and represent the joint angle data and joint load data of the th sampling point respectively, represents the electromyogram signal data of the th related muscle surface among the th sampling point, represents the total number of muscles associated with the target joint, represents the total number of sampling points; In this embodiment, the method for obtaining the low-dimensional latent variable representation sequence is: Input the joint feature vector of each sampling point into the trained variational autoencoder, output the mean vector and logarithmic variance vector of the latent variable, and sample the corresponding latent variable representation of each sampling point from the latent variable distribution based on the reparameterization technique. The formula is: where and represent the latent variable representation and mean vector of the th sampling point respectively, , represents the logarithmic variance vector, is the standard normal noise, that is obeys the multi-dimensional Gaussian distribution with a mean of 0 and a covariance matrix of the identity matrix; Form a low-dimensional latent variable representation sequence according to the time stamps corresponding to the sampling points. The latent variable representation sequence is .

[0050] In this embodiment, in order to perform non-linear compression on the collected high-dimensional multi-source joint feature vectors, extract the core characterization features of the action behavior, and at the same time retain the joint expression relationship between the joint angle, electromyogram signal and load information, the system uses a variational autoencoder (VAE) model to extract the low-dimensional latent variable representation. Through this model, the joint feature vector at each sampling point can be mapped to a low-dimensional latent variable vector with continuity and distribution characteristics, forming a latent variable representation sequence for subsequent time modeling and anomaly analysis.

[0051] Specifically, for each sampling point , the collected joint feature vector includes the angular data of the target joint , the RMS values of the EMG signals of multiple muscles associated with the joint and the corresponding joint load data . This vector serves as the input to the variational autoencoder model, which consists of an encoder network and a decoder network. The encoder is used to compress high-dimensional features into latent variable distribution parameters, and the decoder is used to reconstruct the original input for supervising model training.

[0052] The encoder network first receives the input joint feature vector , and through several layers of fully connected neural network structures, outputs two vectors: one is the mean vector of the latent variable distribution , and the other is the log variance vector .

[0053] Since directly sampling from a probability distribution is non-differentiable and cannot be used for backpropagation training of neural networks, this embodiment uses the reparameterization technique to achieve differentiable sampling. The reparameterization method transforms the random sampling process into a differentiable transformation of model learnable parameters and independent noise variables, specifically: where is an independent Gaussian noise vector sampled from a multi-dimensional standard normal distribution. Through the above transformation, the latent variable representation corresponding to each sampling point is obtained , and this latent variable can be optimized through gradient backpropagation during training.

[0054] The decoder network structure is symmetric to the encoder, with its input being the latent variable , and the output being the reconstructed joint feature vector , and the reconstruction error is calculated with the true input . The training objective of the entire variational autoencoder is to minimize two parts of the loss function: one is the reconstruction loss, which is used to ensure that the latent variable has the ability to preserve fidelity and can recover the original multi-channel features; the other is the KL divergence loss, which is used to approximate the latent variable distribution to the standard normal distribution, thereby obtaining an encoding space with good distribution structure and continuity.

[0055] The overall training objective function of the variational autoencoder is: where is the reconstruction error, and the second term is the KL divergence, which is used to measure the distance between the latent variable distribution and the standard normal distribution. The model uses the Adam optimizer to update parameters. The training data is the sequence of joint feature vectors of all sampling points, and the training process is completed through repeated iterations of batch input, forward propagation, loss calculation, and backpropagation.

[0056] is the variational distribution, representing the conditional probability distribution of the latent variable given the input , which is calculated through the encoder network. Its output is the mean and log variance of the latent variable, used to construct the Gaussian distribution of the latent variable. , which is used to construct the Gaussian distribution of the latent variable. is the reconstruction error, representing the difference between the data reconstructed by the model through the latent variable and the real input data . Here, the squared Euclidean distance is used, that is, the sum of the squares of the errors of each feature. The goal of the reconstruction error is to minimize this difference so that the input can be effectively reconstructed.

[0057] represents the expectation of the latent variable . Since is obtained from the probability distribution sampled from , the average value of the reconstruction error is obtained by calculating the expectation through sampling all latent variables. This part reflects the quality of the model's reconstruction ability. The reconstruction error measures whether the decoder can successfully convert the latent variable back to the input data after the input is mapped to the latent space.

[0058] The prior distribution of the latent variable is a standard normal distribution, that is, a Gaussian distribution with a mean of 0 and a variance of 1. The prior distribution reflects the distribution that the latent variable should follow when the input data is not observed. is the KL divergence, which is used to measure the difference between the variational distribution and the prior distribution . The smaller the KL divergence, the closer the variational distribution is to the prior distribution. The goal of the KL divergence is to make the variational distribution as close as possible to the standard normal distribution, thereby ensuring that the latent space has a good distribution structure and avoiding unreasonable distributions of latent variables during the training process.

[0059] Finally, the trained variational autoencoder model can be used to infer the newly input joint feature vectors. The system inputs the joint feature vectors at each time point into the trained encoder, outputs the corresponding latent variable mean and variance, and obtains the low-dimensional latent variable representation of each sampling point based on reparameterization sampling. The latent variables of all sampling points form a latent variable representation sequence in chronological order, and this sequence will serve as the input basis for the subsequent time series prediction model.

[0060] The method for predicting the latent variable representation of the next sampling point based on the latent variable representation of the previous sampling point in the latent variable representation sequence is as follows: Use the trained gated recurrent unit to process the latent variable sequence for time series modeling. The latent variable representation of each sampling point is used as the input. The initial hidden state of the gated recurrent unit is 0. It receives the latent variable representation of the current sampling point and the hidden state of the previous sampling point, and calculates the hidden state of the current sampling point through the gating mechanism; Based on the hidden state of the current sampling point, use a linear layer to predict the latent variable representation of the next sampling point. The logic is as follows: Among them, represents the predicted latent variable representation of the th sampling point, represents the hidden state of the th sampling point, and represent the learned weights and biases of the gated recurrent unit respectively.

[0061] In this embodiment, to further model the dynamic evolution law of individual joint movement behaviors in the time dimension, the system performs time series modeling on the latent variable representation sequence output by the variational autoencoder and constructs a latent variable trajectory prediction model using a gated recurrent unit (GRU). The model is used to capture the dependencies of latent variables in the time series and realize the prediction of the latent variable state at future times to assist subsequent anomaly detection and risk analysis.

[0062] GRU is a recurrent neural network structure with short-term memory ability. Its feature is to selectively remember and forget information through the gating mechanism, so it is more efficient and stable when processing time series data. This model accepts each time step in the latent variable sequence as the input, and combines the historical hidden state to output the updated hidden state at the current moment for passing information to the next moment. The first hidden state of the GRU network is a zero vector, ensuring that the model starts behavior modeling from an empty state.

[0063] At each time step, the GRU network calculates the intermediate candidate state and the final hidden state in sequence according to the current input latent variable and the hidden state of the previous time step through its internal update gate and reset gate mechanisms. The update gate determines how much historical information should be retained in the current state, and the reset gate determines the degree of dependence between the current input and the historical state, thereby realizing the modeling of the dynamic evolution trend of the latent variable. This hidden state can be regarded as an abstract encoding of the change trend of the latent variable in the previous time step. Based on the hidden state at the current moment, predictions are made through a set of linear transformation layers (i.e., fully connected layers), and the latent variable representation of the next time step is output.

[0064] In the training phase, to ensure that the latent variable representation has effective application significance in subsequent anomaly detection and risk assessment steps, only motion data within the healthy range is used during the training process. Specifically, multi-channel data such as joint angles, electromyogram signals, and joint loads are collected from the motion process of healthy individuals, and these high-dimensional data are mapped to the latent variable space through a variational autoencoder (VAE). During training, only motion data that conforms to the normal physiological range of the human body is selected to ensure that the latent variable representation learned by the model conforms to the physiological pattern of normal behavior. Ensure that the selected training data only comes from healthy individuals during training and exclude any samples of pathological or abnormal movements. For example, for joint angle data, the system only collects angles within the normal motion range, such as the joint movement range during normal walking, running, and standing. These data will be used as training data and input into the VAE model. After passing through the encoder network, the latent variable representation is obtained, and the normal physiological patterns of movement amplitude, muscle activation pattern, and load change are captured in the latent variable space.

[0065] The encoder of the variational autoencoder (VAE) model maps the joint feature vector of each sampling point to the latent variable space to generate latent variables, which represent the motion state of the target joint within the normal activity range. To ensure that the distribution of the latent variable space conforms to the normal physiological range, the system can normalize the data by normalizing each feature of the input data to a specific interval, thereby realizing the normalization of the latent variable. The variational autoencoder training method of this embodiment ensures that the model can only learn the motion trajectories within the normal joint activity range during the training process, which is of great significance for subsequent anomaly detection and risk warning. In this way, the system can effectively distinguish normal and abnormal behaviors and identify abnormal transitions in a timely manner, thereby providing accurate risk assessment and warning for individuals.

[0066] The GRU network structure mainly includes two core gating units: the update gate and the reset gate. The update gate controls how much of the previous moment's memory is retained in the current hidden state, and the reset gate controls the degree of integration of the current input with the past state. This mechanism can suppress the vanishing gradient problem when dealing with complex non-linear time series, improve the model's ability to model long-term dependence behavior, and thus enhance the stability and accuracy of latent variable trajectory prediction.

[0067] The role of the update gate is to control the update degree of the hidden state at the current moment. Specifically, the update gate determines how much of the historical memory should be retained in the current hidden state and to what extent it depends on the input information at the current moment. If the output of the update gate is large, it means that the information at the current moment should retain more of the hidden state from the previous moment; conversely, it means that more dependence should be placed on the input data at the current moment. In this way, the update gate allows the network to flexibly adjust the dependence on past memories and current inputs during the training process.

[0068] The reset gate is used to control the degree of combination of the current input and the hidden state of the previous moment. The reset gate determines the "forgetting" degree between the current input and the historical hidden state. When the output of the reset gate is small, it means that the current hidden state hardly depends on the historical state but focuses more on the current input data; when the output of the reset gate is large, the influence of the historical hidden state on the current hidden state is strong. This mechanism enables the GRU to flexibly adjust the memory of past information and the response to new information when facing non-stationary behaviors or rapid changes in the time series.

[0069] In this embodiment, the latent variable representation at each sampling point is a low-dimensional state representation compressed based on multi-source data such as joint angles, electromyogram signals, and joint loads at that moment. These state representations have a high degree of temporal dependence. Especially during continuous actions, some posture transitions are slow (such as the process of getting up), while some abnormal behaviors will trigger mutations in the latent variables (such as sitting down suddenly or a sudden increase in load). At this time, the gating mechanism will judge whether to retain the hidden state of the previous moment to prevent the model from "misremembering" sudden abnormalities as normal trends.

[0070] After the data at the current sampling point is input into the GRU, the update gate first compares the current latent variable representation with the hidden state information of the previous moment to determine whether the current prediction should rely more on past trends or quickly respond to changes in the current data. For example, when an individual is in a state of continuous standing or slow walking, the update gate tends to retain more historical information to maintain continuity and prediction stability; when an individual suddenly stands up from a squatting position and the electromyogram suddenly increases, the update gate will increase the weight of the current input, enabling the model to quickly perceive and respond to this physiological change.

[0071] The reset gate is mainly responsible for determining whether the current latent variable needs to completely break away from the influence of the historical state during this process, and is particularly applicable to some drastic posture transitions or abnormal transition behaviors. Taking the knee joint as an example, when an individual suddenly jumps up rapidly from a stationary state, accompanied by drastic changes in electromyogram and load, the reset gate will reduce the influence of historical information on the current state, making the prediction of the current latent variable more sensitively reflect the starting stage of the jumping behavior. Finally, through the collaborative action of the update gate and the reset gate, the hidden state maintained by the GRU can adaptively reflect the current physiological state evolution trend of the individual at each sampling point. This mechanism has extremely high practical value for judging whether the behavior evolves smoothly or whether abnormal transitions occur. It ensures that the prediction of the model is not only a linear continuation of time, but also a dynamic modeling of the spatio-temporal changes in complex motion scenarios.

[0072] Step 4: Calculate the deviation between each actual latent variable and the predicted latent variable, determine whether there is an abnormal transition behavior between the multi-channel data corresponding to adjacent sampling points according to the deviation, map the abnormal transition behavior into a directed graph network, count the cumulative frequency and edge weight distribution of the abnormal transition behavior, and divide the risk level, and output the corresponding warning signal for musculoskeletal disorders.

[0073] In this embodiment, the formula for calculating the deviation between the predicted latent variable representation and the true latent variable representation of each sampling point is: where represents the deviation between the predicted latent variable representation and the true latent variable representation of the th sampling point, and respectively represent the predicted latent variable representation and the true latent variable of the th sampling point. When , it is determined that there is an abnormal transition behavior between the multi-channel data of the th sampling point and the th sampling point, represents the deviation threshold.

[0074] In this embodiment, the latent variable representation sequence is extracted from the high-dimensional joint feature vector through a variational autoencoder model, representing the low-dimensional compressed expression of joint angles, electromyogram signals, and load states. This latent variable space has the characteristics of continuity, predictability, and physiological behavior consistency. Therefore, when the system predicts the latent variable state at the next moment, the size of its error actually reflects whether the "current motion state conforms to the existing behavior evolution law".

[0075] During the evolution of normal behavior, the joint movements, muscle activations, and load transmissions of an individual usually exhibit a continuous and gradually changing pattern. The GRU model trained on normal samples can accurately learn the evolution trend of this time series, and thus maintain a low deviation when predicting the next state. If the current state is normal, the model will be able to "predict accurately"; conversely, if there is a sudden change in the current behavior, this change will break the original evolution trend, resulting in the model being unable to accurately predict the next latent variable, and the error will increase accordingly. When the prediction deviation of the latent variable exceeds the threshold, it indicates that the current behavior is beyond the evolution range that the model "considers normal", that is, this behavior is unpredictable for the model. Since the model is trained based on healthy, continuous, and regular behavioral data, its "understanding" is itself based on normal samples, so its prediction error can naturally be used as an indicator for judging "anomaly".

[0076] The change of the latent variable represents the joint behavioral state of multi-channel physiological data (angle + electromyogram + load) in the compressed space. Sudden changes in joint angles, strong muscle electrical activities, or load impacts will all cause "drastic deviations" in the latent variable space, and this kind of deviation can be sensitively recognized by the system through the error judgment mechanism. Therefore, it actually corresponds to "nonlinear jumps in physiological behavior".

[0077] Regarding the setting of the deviation threshold, this embodiment provides two methods: one is the static threshold setting method, that is, based on the prediction error distribution of normal samples in the training set, select its 95% or 99% quantile as the threshold to ensure that the vast majority of normal behaviors will not trigger anomaly judgments; the other is the dynamic threshold adaptation method, that is, combine the individual historical data or the mean and standard deviation of the deviation within the sliding window to dynamically adjust the numerical range to improve the individual adaptation ability of anomaly detection. For specific usage scenarios, such as rehabilitation training or industrial operation scenarios, the system can also adopt expert experience to set a fixed threshold to meet the requirements of specific behavior recognition accuracy.

[0078] In this embodiment, the method for outputting the corresponding musculoskeletal disorder warning signal is as follows: Map the abnormal transition behaviors to the directed graph network, obtain the edge weights corresponding to each abnormal transition behavior, count the frequencies of the abnormal transition behaviors, and combine the deviations corresponding to each abnormal transition behavior to calculate the transition risk state of the target joint during the observation process. The formula is as follows: Among them, represents the transition risk state, represents the number of all sampling points, that is, represents the number of all edge weights, and respectively represent the edge weight and deviation corresponding to the th abnormal transition behavior. Represents the total number of all abnormal transition behaviors; If , issue a low - risk warning indicating occasional abnormal transitions with low edge weights. If , issue a medium - risk warning indicating frequent abnormal transitions or significant edge weights. If , issue a high - risk warning indicating multiple abnormal transitions occurring on high - edge - weight paths. If , issue an ultra - high - risk warning requiring immediate intervention. In a typical normal working movement scenario, occasional abnormal transitions (such as single myoelectric imbalances or slight joint - angle abnormalities) may be detected, but due to low edge weights and small deviations, the overall risk score will still remain at a low level. At this time, the RS value usually ranges between 0.05 and 0.15, corresponding to the low - risk level, and is only used for indicating behavioral fluctuations.

[0079] When an individual has multiple abnormal transitions in a short period of time, or some transition paths have relatively high structural edge weights (such as large - angle jumps, sudden myoelectric increases, or severe load fluctuations), the RS value will gradually increase. Generally speaking, in a mildly unstable state, the RS value is concentrated in the range of 0.15 to 0.30, corresponding to the medium - risk level, indicating attention should be paid to the potential abnormal patterns existing in the behavioral evolution. In more serious situations, such as continuous multiple high - intensity abnormal transitions in an individual's behavior, or transitions concentrated in areas where the system has marked high - edge - weight paths (such as deep flexion, standing up under high load, etc.), the RS value will further increase to between 0.30 and 0.50. This numerical range indicates that the system's prediction ability has failed multiple times, and the individual's behavior significantly deviates from the normal trajectory, and the risk level is determined as "high - risk". If the RS value further exceeds 0.50, it indicates that continuously detected abnormal transitions are extremely costly in terms of structure and have severe dynamic deviations, often accompanied by potential physiological damage risks or functional out - of - control behaviors, such as severe falls, continuous dislocations, and severe muscle - group imbalances. At this time, the RS has reached the "ultra - high - risk" threshold, and an early - warning mechanism at the intervention level will be immediately triggered, suggesting aborting the current action and issuing a health - risk reminder.

[0080] In this embodiment, "abnormal transition behavior" is not equivalent to traditional acute trauma events, but refers to potential physiological state mutations caused by factors such as muscle fatigue, neural control imbalance, or postural stability decline during continuous movements. This behavior models multi - source time - series data by combining variational autoencoders and gated recurrent units, and realizes the quantitative analysis of dynamic evolution paths with the help of a standard pose graph network structure, thereby identifying non - linear abnormalities in action patterns. This modeling method effectively captures the inherent characteristics of the chronic accumulation of musculoskeletal disorders, making "abnormal transition behavior" an important indicator for predicting and warning of musculoskeletal disorder risks, and achieving early identification and intervention of potential injuries.

[0081] The transition risk state reflects two key dimensions of risk superposition: the structural burden of the transition itself and the degree of deviation of the behavioral state from normal healthy behavior. The edge weight reflects "how dangerous this path itself is". Each transition from one standard posture to another has a physiological "load intensity". This value is a weighted index composed of angular mutations, myoelectric bursts, and changes in plantar load. A large edge weight indicates that this transition has a high physiological cost to the musculoskeletal system itself. Even if no abnormalities occur, it belongs to a "high-cost action". Therefore it is an innate risk.

[0082] The latent variable deviation reflects how severe the actual behavior deviates from the predicted healthy behavior direction, reflecting the difference between the predicted normal healthy behavior of the model and the actual behavior, meaning that the behavior has deviated significantly and a situation prone to physiological damage has occurred. The combination of the two enables the model to consider both the physiological intensity of the action and the degree of prediction failure when evaluating abnormal behaviors, thus more comprehensively identifying potential risks.

[0083] A true serious risk of musculoskeletal disorders is constituted if and only if both the structural burden and the degree of deviation of the behavioral state from normal healthy behavior occur simultaneously. If only the latent variable deviation is large, but the edge weight of this transition path itself is low, the risk cannot be said to be high. If only the edge weight is high, but the latent variable prediction is accurate, this indicates that although it is a high-cost action, the individual executes it stably and it is not considered a high risk. Only the superposition of the two constitutes the most vigilant situation in abnormal transitions.

[0084] Using a normalization factor is to standardize the risk score so that it is not affected by the monitoring time or data length, which ensures the comparability of scores under different monitoring windows and evaluates the "risk density" per unit time.

[0085] In this embodiment, the transition risk state scoring formula comprehensively considers the physiological burden of the transition action itself (reflected by the edge weight) and the difference between the predicted normal healthy behavior of the model and the actual behavior (reflected by the latent variable deviation) by normalizing the risk contribution value of cumulative abnormal transitions. The product of the two constitutes a single score of the transition risk. After statistically analyzing all abnormal transitions and performing normalization processing, a comparable RS index is formed, which can be used to drive the subsequent four-level risk warning mechanism.

[0086] In a typical monitoring period, assuming the sampling time window is 100 minutes and data is collected once every 30 seconds, the total number of transitions is approximately 200 times. Under normal circumstances, the number of abnormal transitions usually accounts for a very small proportion of the total number of transitions, generally not exceeding 10%. This means that in 100 total transitions, at most 10 abnormal transitions may occur. The number of abnormal transitions is relatively small, and the edge weight of each abnormal transition is determined by the differences in joint angle changes, electromyogram changes, and joint loads. According to physiological principles and movement patterns, the value of the edge weight generally ranges from 0.3 to 1.5. Transitions with larger edge weights usually involve large-scale joint movements or a greater degree of muscle activation. The latent variable prediction error reflects the gap between the model prediction and the actual observation, usually ranging from 0 to 2. An error greater than 1 indicates a relatively large prediction deviation. However, under normal circumstances, the latent variable error is between 0 and 1, and a deviation exceeding 1 usually indicates a very significant abnormal behavior.

[0087] Under normal circumstances, assuming no abnormal transitions occur, that is, the summation term in the formula is 0, the transition risk state will be 0. This means that the system prediction is completely consistent with the actual observation, and no abnormal events occur, indicating a "risk-free" state. Assuming there are extreme abnormal behaviors, these behaviors will result in the maximum edge weight and the maximum latent variable deviation. In this case, there is also a theoretical upper limit for the maximum value of the transition risk state. Assuming the total number of transitions is 100 times and the number of abnormal transitions is 10 times, the edge weight of each abnormal transition takes a relatively large edge weight value, set to 1.5, and the prediction deviation of each latent variable is set to a relatively large prediction error, set to 2. Substituting these values into the formula, the calculated transition risk state is 1.5. Under common settings, the maximum value of the transition risk state is approximately around 0.5. Of course, the probability of extreme abnormal behaviors is very low, so such a high transition risk state rarely occurs in practice.

[0088] In a normal and healthy state, the deviation between the model prediction and the actual behavior is small, and the occurrence frequency of abnormal transitions is also very low. Therefore, the RS value usually concentrates between 0 and 0.15. That is to say, under normal circumstances, the behavior of the system is stable. When relatively frequent abnormal transitions occur in an individual's movement behavior, or these transitions occur on paths with higher joint loads and larger angle changes, the RS value will gradually increase and may reach 0.3 or higher. If the number of abnormal transitions is very large and the prediction deviation is very large, the RS value will increase significantly, but usually it will not exceed 0.5. In extreme abnormal cases, the RS value can exceed 0.5.

[0089] Please refer to Figure 5 , the present invention further provides a risk assessment and early warning system for musculoskeletal disorders based on multi-source data. The system is used to execute the above-mentioned risk assessment and early warning method for musculoskeletal disorders based on multi-source data, including: A feature acquisition module for continuously acquiring multi-channel data of a target joint to form a joint feature vector sequence. The multi-channel data includes joint angle data corresponding to time stamps, electromyogram signal data of the muscle surface associated with the target joint, and corresponding joint load data. The joint feature vector sequence includes a joint angle sequence, an electromyogram signal sequence, and a joint load sequence; A network construction module for performing K-means clustering on the joint angle sequence, dividing all joint angle data into multiple standard postures, constructing edge weights between any two standard postures according to the joint angle data, electromyogram signal data, and corresponding joint load data within each standard posture, and constructing a directed graph network representing the behavior evolution path in the acquisition order; A variable prediction module for inputting the joint feature vector sequence into a variational autoencoder model for feature compression to obtain a low-dimensional latent variable representation sequence, inputting the latent variable representation sequence into a time series prediction model, using a gated recurrent unit for time series modeling, and predicting the latent variable representation of the next sampling point based on the latent variable representation of the previous sampling point in the latent variable representation sequence; A risk assessment module for calculating the deviation between each actual latent variable and the predicted latent variable, determining whether there is an abnormal transition behavior between the multi-channel data corresponding to adjacent sampling points according to the deviation, mapping the abnormal transition behavior into the directed graph network, statistically analyzing the cumulative frequency and edge weight distribution of the abnormal transition behavior, and dividing the risk level, and outputting a corresponding warning signal for musculoskeletal disorders.

[0090] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by software simulation of a large amount of collected data to get a formula closest to the real situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0091] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed by hardware or software methods depends on the specific application and design constraints of the technical solution.

[0092] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units. They can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0093] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application.

Claims

1. A method for risk assessment and early warning of musculoskeletal disorders based on multi-source data, characterized in that The specific steps include: Continuously collect multi-channel data of the target joint to form a joint feature vector sequence. The multi-channel data includes joint angle data corresponding to time stamps, electromyogram signal data on the muscle surface associated with the target joint, and corresponding joint load data. The joint feature vector sequence includes a joint angle sequence, an electromyogram signal sequence, and a joint load sequence; Perform K-means clustering on the joint angle sequence, divide all joint angle data into multiple standard postures, construct the edge weights between any two standard postures according to the joint angle data, electromyogram signal data, and corresponding joint load data within each standard posture, and construct a directed graph network representing the behavior evolution path in the acquisition order; Input the joint feature vector sequence into a variational autoencoder model for feature compression to obtain a low-dimensional latent variable representation sequence. Input the latent variable representation sequence into a time series prediction model, use a gated recurrent unit for time series modeling, and predict the latent variable representation of the next sampling point based on the latent variable representation of the previous sampling point in the latent variable representation sequence; Calculate the deviation between each actual latent variable and the predicted latent variable, determine whether there is an abnormal transition behavior between the multi-channel data corresponding to adjacent sampling points according to the deviation, map the abnormal transition behavior into the directed graph network, statistically analyze the cumulative frequency and edge weight distribution of the abnormal transition behavior, and divide the risk level, and output the corresponding musculoskeletal disorder warning signal.

2. The method for risk assessment and early warning of musculoskeletal disorders based on multi-source data according to claim 1, wherein: The target joint is one of the hip joint, knee joint, or ankle joint in the lower limb joints, and the joint load data is the average pressure on the soles of both feet; When collecting multi-channel data of the target joint, set multiple sampling points, synchronously collect joint angle data, electromyogram signal data, and joint load data at each sampling point, and the time interval between adjacent sampling points is the same.

3. The method for risk assessment and early warning of musculoskeletal disorders based on multi-source data according to claim 1, characterized in that: The logic for performing K-means clustering on the joint angle sequence is: Determine the maximum movement angle of the target joint, divide the maximum movement angle into angle ranges corresponding to K standard postures, and use the median of the angles within each angle range as the initialized clustering center; Calculate the Euclidean distance between each joint angle data and each initialized clustering center, assign the corresponding joint angle data to the nearest clustering center, and update each clustering center to the mean of all angle values in its clustering cluster; Repeat calculating the Euclidean distance between each joint angle data and each updated clustering center, reassign the corresponding joint angle data to the nearest clustering center until the clustering center converges, that is, the clustering center no longer changes; Map the standard posture corresponding to the cluster where each joint angle data is located to the standard posture corresponding to each joint angle data.

4. A method for risk assessment and early warning of musculoskeletal disorders based on multi-source data according to claim 2, characterized in that: The method for constructing the edge weights between any two standard postures is: Calculate the average of the joint angles within the clustering clusters corresponding to the two standard postures, the root mean square value of all EMG signals, and the average of the joint loads respectively, and calculate the difference values of the joint angles, EMG signals, and joint loads between the two clustering clusters. The formulas are as follows: Among them, , and respectively represent the difference values of the joint angles, EMG signals, and joint loads between the clustering clusters corresponding to the standard postures of and . and respectively represent the average of the joint angles of the corresponding clustering clusters when the standard postures are and . and respectively represent the average of the joint loads of the corresponding clustering clusters when the standard postures are and . and respectively represent the root mean square value of the EMG signals on the th and th associated muscle surfaces in the corresponding clustering clusters when the standard postures are . and both represent the numbers of the standard postures, and respectively represent the number of the muscle associated with the target joint and the total number of the associated muscles; Weight the difference value to construct the edge weight between two standard postures, based on the formula: Among them, represents the standard posture The edge weight when transitioning to the standard posture respectively represent the angle weight, signal weight and load weight, and .​ 5. A method for risk assessment and early warning of musculoskeletal disorders based on multi-source data according to claim 4, characterized in that: The method for constructing a directed graph network representing the behavior evolution path is: Taking the joint angle data, EMG signal data, and corresponding joint load data at each sampling point as a time node, sorting all time nodes according to the acquisition timestamps corresponding to the acquisition points, the direction of the directed graph network is from the previous time node to the next time node, and the edge weight between two adjacent time nodes is the edge weight when transitioning from the standard posture corresponding to the previous time node to the standard posture corresponding to the next time node.

6. The method for risk assessment and early warning of musculoskeletal disorders based on multi-source data according to claim 2, wherein: The joint feature vector sequence is as follows: Wherein, is the joint feature vector of the th sampling point, and respectively represent the joint angle data and joint load data of the th sampling point, represents the electromyogram signal data of the th associated muscle surface among the th sampling points, represents the total number of muscles associated with the target joint, represents the total number of sampling points; The method for obtaining the low-dimensional latent variable representation sequence is as follows: The joint feature vectors of each sampling point are input into the trained variational autoencoder, and the mean vector and logarithmic variance vector of the latent variables are output. Based on the reparameterization technique, the latent variable representation corresponding to each sampling point is sampled from the latent variable distribution, and the formula is as follows: where and represent the latent variable representation and the mean vector of the -th sampling point respectively, , represents the logarithmic variance vector, is the standard normal noise, that is follows a multi-dimensional Gaussian distribution with a mean of 0 and a covariance matrix of the identity matrix; Form a low-dimensional latent variable representation sequence according to the time stamps corresponding to the sampling points, and the latent variable representation sequence is , and the method for predicting the latent variable representation of the next sampling point based on the latent variable representation of the previous sampling point in the latent variable representation sequence is: Use the trained gated recurrent unit to perform temporal modeling on the latent variable sequence Each latent variable representation at a sampling point is used as input. The initial hidden state of the gated recurrent unit is 0. It receives the latent variable representation at the current sampling point and the hidden state at the previous sampling point, and calculates the hidden state at the current sampling point through the gating mechanism; Based on the hidden state of the current sampling point, use a linear layer to predict the latent variable representation of the next sampling point. The logic is as follows: Among them, represents the predicted latent variable representation of the th sampling point, represents the hidden state of the th sampling point, and respectively represent the learned weights and biases of the gated recurrent unit.

7. A method for risk assessment and early warning of musculoskeletal disorders based on multi-source data according to claim 5, characterized in that: Calculate the deviation between the predicted latent variable representation and the true latent variable representation for each sampling point, based on the formula: where represents the deviation between the predicted latent variable representation and the true latent variable representation for the -th sampling point, and represent the predicted latent variable representation and the true latent variable for the -th sampling point respectively. When , it is determined that there is an abnormal transition behavior between the multi-channel data of the -th sampling point and the -th sampling point. represents the deviation threshold.

8. The method for risk assessment and early warning of musculoskeletal disorders based on multi-source data according to claim 2, characterized in that: The method for outputting the corresponding musculoskeletal disorder warning signal is as follows: Map the abnormal transition behavior to a directed graph network, obtain the edge weights corresponding to each abnormal transition behavior, count the frequency of the abnormal transition behavior, and calculate the transition risk state of the target joint during the observation process in combination with the deviation corresponding to each abnormal transition behavior. The formula is as follows: Among them, represents the transition risk state, represents the number of all sampling points, that is, it represents the number of all edge weights, and respectively represent the edge weight and deviation corresponding to the th abnormal transition behavior, represents the total number of all abnormal transition behaviors; If , issue a low-risk warning indicating an occasional abnormal transition with low edge weight. If , issue a medium-risk warning indicating frequent abnormal transitions or significant edge weight. If , issue a high-risk warning indicating multiple abnormal transitions occurring on a high-edge-weight path. If , issue an ultra-high-risk warning requiring immediate intervention.

9. A risk assessment and early warning system for musculoskeletal disorders based on multi-source data, characterized in that: The system is used to execute a method for risk assessment and warning of musculoskeletal disorders based on multi-source data according to any one of claims 1-8, including: A feature acquisition module, configured to continuously acquire multi-channel data of a target joint to form a joint feature vector sequence. The multi-channel data includes joint angle data corresponding to timestamps, EMG signal data on the surface of muscles associated with the target joint, and corresponding joint load data. The joint feature vector sequence includes a joint angle sequence, an EMG signal sequence, and a joint load sequence; A network construction module, configured to perform K-means clustering on the joint angle sequence, divide all joint angle data into multiple types of standard postures, construct the edge weights between any two standard postures according to the joint angle data, EMG signal data, and corresponding joint load data within each type of standard posture, and construct a directed graph network representing the behavior evolution path according to the acquisition order; A variable prediction module, configured to input the joint feature vector sequence into a variational autoencoder model for feature compression to obtain a low-dimensional latent variable representation sequence, input the latent variable representation sequence into a time series prediction model, perform time series modeling using a gated recurrent unit, and predict the latent variable representation of the next sampling point based on the latent variable representation of the previous sampling point in the latent variable representation sequence; A risk assessment module, configured to calculate the deviation between each actual latent variable and the predicted latent variable, determine whether there is an abnormal transition behavior between the multi-channel data corresponding to adjacent sampling points according to the deviation, map the abnormal transition behavior into the directed graph network, count the cumulative frequency and edge weight distribution of the abnormal transition behavior, divide the risk level, and output the corresponding musculoskeletal disorder warning signal.

Images